A Locally Weighted Fixation Density-Based Metric for Assessing the Quality of Visual Saliency Predictions

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Cognitive Science

Background:

• Computational visual attention (VA) models are increasingly prevalent.
• Traditional evaluation relies on metrics comparing predicted saliency to human eye-tracking data.
• Existing evaluation metrics have significant shortcomings.

Purpose of the Study:

• To identify and illustrate flaws in current visual saliency evaluation metrics.
• To propose a novel evaluation metric for visual saliency prediction.
• To establish a benchmark for assessing saliency prediction quality.

Main Methods:

• Developed a new metric using local weights based on fixation density.
• Created a ground-truth subjective database with human evaluations of 17 VA models.
• Collected ratings from 16 human observers on a five-point scale.
• Correlated metric scores with human subjective ratings.

Main Results:

• The proposed metric significantly outperforms existing popular metrics.
• Human subjective ratings validate the superiority of the new metric.
• The constructed database serves as a valuable benchmark for future research.

Conclusions:

• The proposed local-weighted fixation density metric is a more effective tool for evaluating visual saliency models.
• The new database and methodology provide a robust framework for future saliency model assessment.
• This work advances the field of computational visual attention by improving evaluation standards.